The Saturation option was unbalanced, giving more weight to the color
red and a TON more weight to the color blue. This patch balances the
colors to provide a more even increase in saturation across the RGB
channels.
Closesjp9000/obs-studio#739
When using this filter in/on OpenGL situations can cause crashes with an
"int to vec4" cast. This fix should resolve that issue.
Closesjp9000/obs-studio#739
Using and creating a window can use issues in game capture if multiple
game captures are active, so revert back to using a mutex, and just
ignore the keepalive check failure if injected inside a UWP program
(only check to see if GetLastError reports that it's not found -- if it
returns access denied or any other error, assume it's in a UWP program,
and ignore the keepalive check).
If the backbuffer count is larger than 3, it could still try to assign
backbuffers to pointers beyond the variable's array size when calling
swap->GetBuffer.
Use a d3d9 device and allocator to encode in QSV.
This fixes a random crash that could only happen on Windows 7. The QSV
Deviced returned a DEVICE_FAILURE after a random amount of time with the
old method.
This fix is totally based on Shinck's QSVHelper.exe patch for OBS
Classic (see
https://obsproject.com/forum/threads/0-633b-qsvhelper-exe-was-killed-encode-failed.19230/page-3#post-161984
for more information)
This is more like a proof of concept, but that fix is currently stable
and tested more than 50 hours, with a single session of +14 hours.
That commit doesn't respect all OBS Guidelines. It is currently
recommended to wait for a more "cleaner" implementation.
Due to using FindWindowEx to search for certain windows, certain windows
will show up that aren't usable/capturable. Prevent these windows from
showing up in the window lists.
For some unknown reason, GetWindow will not traverse the entire window
tree. It could be due to Microsoft purposely hiding certain UWP
windows, though the reason is unknown. For some equally unknown reason
FindWindowEx does work in its place.
This fixes the issue of not being able to find/capture certain windows,
such as halo 5: forge.
D3D12 capture does not take in to account when multiple backbuffers are
in used. With previous versions of Direct3D there was no need to do
this, but with D3D12 you must explicitly capture each specific
backbuffer currently in use.
The encoder name was changed from "nvenc_h264" to "h264_nvenc", and will
throw a warning in the log file if you use the former, so try the latter
first, then the former.
Similar to the parent commit, do not reset the blend state if the source
is not marked as a custom drawn source. Filters applied won't have the
correct blend state and the alpha edges of the text will be partially
black.
The blend state does not need to (and should not) be changed. This
fixes an issue where if filters are applied to an image source, the
image source will paint incorrectly.
Replaces the "Color Correction" filter with a newer version that uses a
matrix, adds hue and saturation, and improves the contrast option.
Closesjp9000/obs-studio#708
If capturing a UWP window, do not fall back to matching windows with the
same window class if the exact window is not found, as this will get any
other UWP window on the system (due to the fact that they all have the
same window class name).
Because the hook cannot get the window handle of UWP windows, fall back
to using the window handle stored in the game capture source itself if
it's unavailable from the hook.
Under certain circumstances, the program may not be able to acquire the
window thread ID for a UWP process, but will be able to acquire the
process ID. In this case, it should soft fail and retry, rather than
assume it's unacquirable and stop trying to reacquire.
The "attempting to hook [executable]" message would not display the
correct executable if it's fallen back to a different window with the
same window class.
Now that the game capture hook creates and controls all the
synchronization objects, it's no longer possible to expect that the hook
is fully loaded by the time game capture tries to initialize it. In
that case, allow game capture to retry the hook for a few frames before
assuming something failed.
This detects whether the target process is a UWP process, and then uses
the open_app_* functions for mutexes/events/mapping. Also slightly
refactors named object open functions.
The only way to open named kernel objects within a UWP "app" is to get
the AppContainer SID, and then open the objects with their full
system namespace names via undocumented NT functions.
All named objects (including file mapped shared memory) need to be
created within the hook itself due to the fact that UWP programs cannot
access named objects outside of the UWP process.
Because shared memory needs to be created within the hook, the capture
loop cannot start until the shared memory has been filled with valid
data. Creating an additional "initialize" event fixes this issue.
Additionally, changed the way that named kernel objects are
opened/created. Before, there were functions that would first try to
open named objects and then implicitly create them if opening failed
(assuming that if the hook didn't create it first, game capture would),
now it's been changed so that you can only either explicitly open or
create.
To check to make sure game capture is still active in the capture
program, it currently uses a named event, and then it checks to see if
that named event exists. However with UWP programs, you can't open a
named event outside of the UWP process. FindWindow on the other hand
does work, so instead of checking to see if a named kernel object
exists, create a window and check to see if that window exists.
The "main" windows detected for UWP programs are basically to help
sandbox the programs -- they run in the ApplicationFrameHost process and
help reduce the possibility of other programs trying to access the
actual process window, which is a child window.
To bypass this, go through the list of child windows for the
ApplicationFrameHost window, and then find the one that's attached to
a different process; that different process will always be the target,
and will allows us to open the actual process of the UWP program.
Instead of calling OpenThread, use DuplicateHandle with the minimal
access rights (SYNCHRONIZE) to be able to get the current thread handle
within a UWP program.
This prevents issues with using standard C functions, where microsoft
would otherwise spit out pointless warnings to encourage using
microsoft-specific functions instead.
This reverts commit 0edaebe192.
The commit was actually less optimal than it was before. Shaders are
always fully unfolded and both sides of a branch are always executed.
Despite that fact, the "avoid branching" commit actually *added* two
extra unnecessary instructions with the same number of actual sample
instructions, making it arguably less optimal than before.
This was tested via the D3DDissamble function.
(Note: This commit also modifies the ipc-util/seg-service modules)
When compiling the final project, always compile
ipc-util/get-graphics-offsets/graphics-hook/inject-helper/seg-service
with static MSVC runtimes to prevent the need of requiring the MSVC
runtimes for both architectures.
QSV frame priority bits are not being marked correctly; their priorities
are always marked either low or disposable, whereas I-frames should be
marked as highest (3), P-frames should be marked as high (2), and other
frames should be marked as either low or disposable. This helps ensure
that the correct frames are dropped when frames need to be dropped.
On certain switchable laptops with AMD graphics, due to a driver
bug/issue, the program will crash when querying capabilities. The
commit this updates to fixes that by not querying capabilities on
startup.
There are two variables used for 'priority', the general marked frame
priority, and the drop priority. They can sometimes be different
because some encoders don't always use the correct frame priority.
Whenever a cursor is captured and the cursor icon changes, it creates a
new texture. This isn't particularly optimal, so instead just store a
cache of cursor textures (based on size), and make the textures dynamic.
Doing this will prevent unnecessary texture reallocation.
Branching is not recommended in shaders and only recently gained
widespread support without causing massive branch misprediction issues. Also sampling inside a branch will cause cache misses and additional texture load instructions.
For some reason librtmp treats a close message as a request to reconnect.
It does this syncronously, and uses the same event processing so that
another close request will continue the cycle until the stack is exhausted
and the application crashes.
Fixes https://obsproject.com/mantis/view.php?id=634
Twitch no longer recommend Main profile on their "Broadcast requirements"
page. The number of devices that can't decode High profile is very small
and limited to old, legacy devices such as the 1st generation iPad (six
years old). Additionally, these devices have limited resolution / FPS
support, so streams are often encoded at levels beyond what they could
decode even disregarding the H264 profile requirement. It's not worth
degrading the quality of streams for modern devices just to support these
legacy decoders. Twitch still provides low-complexity transcodes for
devices which cannot handle Source quality streams.
When a pulseaudio source is created from existing settings and the
device is not found, then the private data will return NULL, making it
impossible to change the source to use a different device. NULL should
never be returned when possible, otherwise settings cannot be changed
and properties cannot be displayed.
Closesjp9000/obs-studio#604
The maximum shutdown timeout value was added as a setting, but never
actually fully implemented. This implements it properly, and sets its
default timeout value to 30 seconds.
This commit fixes the issue outlined in the following thread:
https://obsproject.com/forum/threads/50045/
When the image source file was replaced by an outside process, it would
only reload when the file's new timestamp was newer than the file's
previous timestamp. This fixes that behavior. Now an image source
will reload any time the file's new timestamp is different than its
previous timestamp.
Cephas Reis
Jim
Michael Fabian Dirks
derrod
Colin Edwards
Gol-D-Ace
Iblis Lin
Alexandre Biny
shiina424
Richard Stanway
Andrew Surzhynskyi
Ryan Foster
Ricardo Constantino
Warren Turkal
Michael Fabian Dirks
Michael Fabian Dirks
CoDEmanX
Robin Hielscher
Christoph Hohmann
Michael Hoang
hwdro
Andrei Nistor
rpslack